Creator Tips

Trending Topics turns your bot into a co-creator on what's happening right now. Cantina surfaces current names, hashtags, and conversation starters in Imagine so you can drop one in and watch your bot riff on it, which is great for telling today's stories in your bot's voice, exploring culture-of-the-moment scenes, and finding a spark when your brain blanks.

Where to find it

1. Tap Imagine in the bottom nav.
2. Pick a bot — swipe through your bots, or use the search to find one.
3. Tap the text box at the bottom of the screen. This is the step that unlocks Suggested Topics — a live grid of currently trending names, hashtags, and references appears above the keyboard.

One of the most useful workflows on Cantina isn't obvious from the surface: instead of jumping straight into an Imagine video, prompt several selfies first — then use those selfies as the foundation for each scene in a multi-scene video.

The selfie locks in the look (wardrobe, setting, lighting, vibe, camera feel). Once you're happy with it, the video builds motion on top of a look you've already chosen. Each selfie becomes a scene anchor. The video built from those anchors inherits everything you put into them.

This article walks through the workflow, why it works, and patterns for keeping multi-scene videos feeling like one piece.

Why this works

Think of it as storyboarding your video before you build it. You prompt a selfie, see how it looks, tweak the prompt, see the new selfie, and iterate until each starting image is exactly right. Then you build the video from those starting points.

This shifts where you do the visual decision-making. Instead of describing wardrobe, setting, lighting, color, and motion all at once in a single video prompt, you nail the look one scene at a time in still images. Then you bring those images to life in the video.

• Generic selfie → generic scene.
• Layered selfie → layered scene.

If your selfie prompt is rich (subject + action + setting + wardrobe + camera + lighting + color + finish), the scene built from it inherits every layer. See Structure Your Prompt for the recipe.

The workflow

Step 1. Plan your scenes. Sketch a quick arc — morning → noon → night, before → during → after, scene 1 → scene 2 → scene 3. Two to four scenes is a comfortable range.

Step 2. Prompt rich selfies in chat with your bot. Write a layered selfie prompt for each scene — setting, wardrobe, time of day, camera, lighting, color, vibe. (See Starter Prompt Library for examples and Structure Your Prompt for the recipe.)

Step 3. Tweak each selfie prompt until the result looks right. This is where the visual decision-making happens — fast, in still images, before committing to a video.

Step 4. Open Imagine, tap Video, and pick your bot.

Step 5. Open the Add Image picker and tap a selfie. Your recent chat selfies show up in a grid with timestamps ("1m ago", "3m ago", "4m ago"). Tap one to use as the visual anchor for a scene.

Step 6. Imagine generates a dialogue + action prompt for that scene. A "Generating script…" overlay plays while it writes.

Step 7. Land in the Video editor with the scene loaded. Each scene has a Dialogue field (what your bot says) and an Action Prompt field (what your bot does). Edit either or both. Repeat for each scene in your video.

Step 8. Tap Save / Generate to render the final video.

You don't have to do all scenes in one sitting. Build your selfie library over days in chat, then pull from it when you're ready to make a video.

Want a scene or a whole video without dialogue?

Every scene comes with auto-generated dialogue, but you don't have to keep it. Clear the Dialogue field for any scene to drop the dialogue. The Action Prompt still drives motion.

For the full workflow, mix-and-match patterns, and when no-dialogue scenes work best, see No Dialogue Imagine Videos.

A full example — a bot's day

Here's a three-scene arc and the selfie prompts behind it. The screenshots in this section come from a real walkthrough with Carl.

Scene 1 — morning at the desk

At your desk with one cup of fresh coffee, soft morning light, focused expression.

Sets the day. Wardrobe is implied (Carl's standard look), light is soft and morning-warm, vibe is private and focused.

Scene 2 — midday on the street

Walking through a busy city street, midday sun, more energy, warm smile.

Shifts the energy. Light is harsh and bright, the world is full of people, the mood is social and active.

Scene 3 — golden hour at the bar

Winding down at a bar, window seat. Golden hour glow. One mojito on the table.

Lands the day. Light returns to warm and intimate, the mojito gives the scene a prop, the vibe is celebratory and quiet.

Use these three selfies as scene anchors for the Imagine video. The arc carries the viewer through the day even without an explicit story — the time of day, location, and mood do the work.

The final video

Here's what the three selfies become once stitched together as Imagine scenes:

Notice how the arc carries: morning quiet → midday energy → evening wind-down. The wardrobe, light, and mood across the three selfies do most of the storytelling — Imagine builds the motion and dialogue on top.

Keeping your scenes feeling like one video

Multi-scene videos work best when the scenes feel connected. A few ways to do this:

• Repeat one consistent element. Same jewelry, same hairstyle, same color in the wardrobe, same color grade family. One repeated detail makes three scenes feel like one trip.
• Use a color arc. Move through related grades instead of jumping wildly. Warm cream → saturated noon → amber gold all live in the same warm family. Compare to jumping to a cool blue noir scene mid-arc — it'll feel like a different video.
• Match finish cues. If selfie 1 is cinematic film grain and selfie 3 is digital sharp, the scenes will feel like they came from different cameras. Pick one finish and run it through every scene.
• Anchor the camera language too. Three handheld shots feel like one video. One handheld + one static wide + one fisheye feels like three different videos.

Picking your scenes — a few arc patterns

If you're stuck on what scenes to use, here are arcs that work well:

• Time of day. Morning → midday → golden hour → night. Easy to plan, easy to vary lighting.
• Before / during / after. Getting ready → at the event → the morning after.
• Location-to-location. Apartment → walking → destination.
• Day in the life. Wake up → work → dinner → wind down.
• Emotional arc. Calm → chaos → calm. Or anxious → confident.
• Wardrobe change. Same place, different outfit per scene — lets the wardrobe carry the story.

When to use this workflow vs. a single Imagine prompt

Use scene anchors when:

• You want a multi-scene video with a clear arc.
• You want each scene's look to be deliberate (specific wardrobe, lighting, color per scene).
• You're telling a story or building a vignette.
• You want to share the result externally and care about polish.

Skip scene anchors and prompt Imagine directly when:

• You want one short clip, not a multi-scene video.
• You're experimenting and don't have a specific look in mind.
• You're going for casual / chat-side energy.
• You're in a hurry. Direct Imagine is faster.

See Fast Videos vs Imagine Videos for more on when each video tool is the right fit.

When scenes don't feel connected

If your finished video feels like three different videos stitched together, the fix is usually in the selfie prompts — not the video build.

1. Compare the three selfie prompts side by side. Find the layers that disagree (one scene is cinematic film grain, another is digital sharp; one is warm tones, another is cool).
2. Rewrite the outlier to match the family. Pick one of the three as the anchor and pull the others toward it.
3. Re-generate the selfies and rebuild the video.

Small edits to the selfie prompts ripple through to the scenes — you don't have to rebuild from scratch.

Keep going

• Prompting 101 — the basics: what a prompt is and where you'll use them.
• Structure Your Prompt — the eight-layer recipe behind a great prompt.
• Starter Prompt Library — copy-and-remix example prompts across every Cantina surface.
• Fast Videos vs Imagine Videos — which video tool to use for which job.
• Using Imagine to Create Videos — the step-by-step for generating an Imagine video.

Want your bot to just do the thing — no voiceover, no dialogue? You can make any Imagine video silent, or mix silent and spoken scenes inside a single video.

The quick version

1. Open your Imagine video in the editor.
2. Tap Edit Script at the bottom.
3. Open the Dialogue field for the scene you want silent.
4. Delete the text.
5. Tap the check mark to save.
6. Generate the video.

The scene plays silent. The action prompt still drives what your bot does.

How it works

Every Imagine scene has two fields:

• Dialogue — what your bot says
• Action Prompt — what your bot does (and any ambient sound you want)

Both are auto-generated when Imagine writes the script for a new scene. They work independently:

• Delete the Dialogue → the scene plays silent.
• Keep the Action Prompt → your bot still moves through the scene the way you described.

The Action Prompt is what gives your bot direction — keep it in place. You can always edit it to refine what your bot does.

Add ambient sound if you want it

A scene with empty Dialogue plays without a voiceover. If you want ambient sound — rain on a window, café chatter, soft music — describe it in the Action Prompt below the dialogue window. The action prompt drives both motion and the soundscape.

Examples of ambient sound in an action prompt:

• "Carl sips his coffee while soft rain patters on the window outside."
• "Carl walks through the busy market, vendors calling out and city traffic in the background."
• "Carl leans back at the bar, soft lounge music playing under the clink of glasses."

Mix silent and spoken scenes in the same video

Each scene in a multi-scene video is independent. You can:

• Keep dialogue on Scene 1, go silent on Scene 2, more dialogue on Scene 3.
• Open with a silent moment, end with a voiceover.
• Make the whole video silent.
• Whatever the story calls for.

Clear the dialogue per scene. Save each. Generate.

When to go silent

Silent scenes work well for:

• Establishing shots — a slow camera move across a setting before your bot speaks.
• Reaction beats — a look, a gesture, a held expression that's stronger without words.
• Mood pieces — atmosphere-driven scenes where the visuals carry everything.
• Soundtrack-led videos — if you're planning to add music later and don't want a voiceover competing.
• Just because — some videos land harder without speech.

Examples for you to try

Open the Action Prompt for the scene you want, swap in your bot's name (or just describe what's happening), and clear the Dialogue field.

Quiet moment:

[Your bot] walks slowly across an empty room, pauses at the window, looks out at the rain.

Reaction beat:

[Your bot] reads the letter, jaw tightens, eyes drop to the floor.

Mood piece with ambient sound:

[Your bot] sips coffee at a quiet café, soft rain patters on the window outside, distant chatter in the background.

Establishing shot:

Wide view of a sunlit kitchen — [Your bot] steps in, sets down a grocery bag, glances around.

Drop any of these into the Action Prompt, leave the Dialogue blank, and generate.

Keep going

• Selfies as Scene Anchors — the multi-scene video workflow that uses chat selfies as the visual foundation for each scene.
• Action Prompting — How to write the part of your prompt that drives motion, camera moves, and transformations.
• Using Imagine to Create Videos — step-by-step for creating an Imagine video.
• Captions on your video — for the scenes where your bot does speak, customize the caption font, size, or color.
• Fast Videos vs Imagine Videos — which video tool to use for which job.
• Cantina Glossary — quick reference for product terms.

The Action Prompt is the cinematic engine of every video scene you build in Cantina. It drives motion, camera moves, cuts, transformations. Everything that happens on screen, with or without dialogue. This article walks through how to write one, what to layer in, and how to tweak until the scene lands.

What's in your toolkit

The Action Prompt sits in the video editor, just below the Dialogue field. It's where you describe what your bot does and how the camera moves.

A few things you can layer into an Action Prompt:

• Action verbs — struts, glares, transforms, leans, pauses, turns, laughs, walks, looks, reaches, falls
• World context — "This is a high fashion show," "It's pouring rain," "The crowd cheers," "Smoke fills the air"
• Cuts — "Cut to a close up of her face," "Cut to a wide shot," "Cut to her hands"
• Focus shifts — "Focus on her legs," "Focus on the mug," "Focus on the door"
• Camera moves — "Pull back," "The camera tracks forward," "The camera orbits her," "Push in," "Tilt up"
• Transformations — "transforms into a cloud of bats," "morphs into a shadow," "shifts into a wolf"
• Sound cues — "soft rain patters on the window," "vendors calling out in the background," "a distant clock ticks," "thunder rolls overhead"

You don't need all seven in every prompt. Pick the ones that serve the scene.

How to write an Action Prompt

Build your prompt in this order before you generate. The more complete it is up front, the closer the first result will land.

Step 1. Pick your action. Two or three verbs that describe what your bot does — struts, glares, transforms.

Step 2. Ground the world. One sentence of context that anchors the scene — "This is a high fashion show," "It's pouring rain outside."

Step 3. Direct the camera. Add the cuts, focus shifts, and camera moves you want — "Cut to a close up of her face," "Pull back," "The camera tracks forward."

Step 4. Layer in transformations or sound if they serve the scene.

Step 5. Save and generate.

Watch a prompt grow

Same scene, three levels of layering. Each level adds one more layer from the toolkit on top of the last. See how the picture sharpens as the prompt fills out.

The scene: a fashion-show runway moment where the model dissolves into a swarm of bats flying at the camera.

Level 1 — core action

The pale alien woman struts, glares, and transforms into a cloud of bats that fly towards camera.

The verbs do most of the work. Struts sets the motion, glares layers in attitude, transforms signals the change, fly towards camera gives the bats direction. One sentence, one beat.

Level 2 — add world context

The pale alien woman struts, glares, and transforms into a cloud of bats that fly towards camera. This is a high fashion show.

The second sentence grounds the world — a runway, an audience, theatrical lighting. The result inherits the visual language of "fashion show" without you having to spell it all out.

Level 3 — add cuts, focus, and camera moves

The pale alien woman struts, glares, and transforms into a cloud of bats that fly towards camera. This is a high fashion show. Cut to a close up of her face with fangs showing. Cut to a wide shot. Focus on her legs. Pull back. The camera tracks forward, focusing on her.

Now it's a cinematic mini-sequence. Multiple shots inside one scene, a focus shift down to the legs, two camera moves (pull back, track forward). 267 / 500 characters. Still room to layer more.

And the result: 

Pair with empty dialogue for cinematic-only moments

Action Prompts can carry an entire scene without dialogue. Clear the Dialogue field, write a rich Action Prompt, and you get a cinematic moment with no voiceover. The visuals and camera tell the story. See No Dialogue Imagine Videos for the full pattern.

Add a narrator

You can write a narrator's voiceover into the Action Prompt itself, alongside what the camera sees and what your bot says. Structure it like a mini screenplay. Three blocks, separated by blank lines, each block labeled for the voice that carries it.

A grumpy blue water bottle sits inside a fridge, staring through the glass shelf with his usual annoyed expression.

NARRATOR: "Meet irritated hydration, the grumpiest bottle in the fridge."

DIALOGUE (irritated hydration lip sync): "Can somebody close the fridge already."

What each block does:

• The action block describes what's on screen — no voice tag, no dialogue.
• NARRATOR is the voiceover. No character on screen speaks this line. It plays over the visuals.
• DIALOGUE (character lip sync) is your bot's spoken line. The lip sync tag tells the model the character's mouth should move to it.

Use a narrator when you want to introduce a character, set up a joke, or land a closing line that the character themselves wouldn't say. It also lets you keep the bot's own dialogue short and reactive. The narrator carries the exposition, the bot carries the attitude and action.

If the result isn't quite right

Before you generate again, look at your prompt and name the missing layer. Most off results trace back to one of these:

• Action looks flat? Stack one or two more verbs.
• World feels generic? Add a sentence of context.
• Camera stays on one angle too long? Add a cut or a focus shift.
• Shot lacks character? Add a transformation, an attitude verb, or a sound cue.

Name the missing layer, add it, then generate.

Keep going

• Structure Your Prompt — the 8-layer recipe behind a great prompt.
• No Dialogue Imagine Videos — how empty dialogue plus a rich Action Prompt makes a cinematic-only scene.
• Selfies as Scene Anchors — multi-scene videos built from chat selfies.
• Starter Prompt Library — copy-and-remix example prompts across every Cantina surface.
• Prompting 101 — the basics.

Got an Action Prompt to share?

Found a pattern that consistently lands? Share it with the community in The Bot Place. We add new patterns over time.

It's a bot, until you add character.

You've created your bot. Now let's explore more prompting tools.

You've created your bot. You've worked through Prompting 101 for the foundations. You may have borrowed a few templates from the Starter Prompt Library.

This article helps you add more depth to your character. You'll find tools for adding layers, writing strong well-rounded prompts, bringing a flat bot to life, and directing the action when your bot makes videos.

We'll work through it in two parts. Part 1 sharpens the personality. Part 2 directs the action.

Start with an archetype

A named archetype gives you a strong starting silhouette. Before you write a single line of prompt, outline what type of character you're making. Borrow from known tropes or archetypes, or create your own. A few from the Starter Prompt Library:

• Roast Bot — sharp, observational, doesn't soften the burn.
• Mentor Bot — warm, patient, structured.
• Disappointed Dad — emotionally restrained, devastating with a single sigh.

Each archetype implies a voice, a posture, a comedic or emotional register.

Add layers to your bot

A bot becomes a character when you give it dimensions beyond a single label. Bot Building Basics covers the three pillars — appearance, voice, and personality. This article gets specific about how to add depth to personality.

Layer a few lines in your prompt:

Identity — who they are, what they want, what they're for.

You are a retired stand-up comic turned line cook. You write recipes the way you used to write jokes.

Voice — how they speak. Rhythm, vocabulary, signature phrases.

You speak in short, punchy sentences. You drop articles like "the" when you're irritated. You say "listen" before making a point.

Quirks — specific tics, hangups, recurring themes.

You bring up your ex-wife unprompted. You hate cilantro and will say so. You always end recipes with "don't overthink it."

Constraints — what they refuse to do or never say.

You never apologize for being blunt. You don't soften criticism.

Escape hatches — what they fall back to when they don't know.

If asked about something outside cooking, deflect with a one-liner and steer back to food.

Once you define each line, you can stack these as the bot's identity.

Tell them what not to do

Negative prompts are as powerful as positive ones. Telling a Roast Bot don't soften the burn is more useful than just describing the burn. Telling a Mentor Bot never use sarcasm keeps their tone encouraging and light, instead of condescending or unkind.

A strong prompt isn't all negative. Use a few well-placed don'ts to keep your character within the boundaries you've set.

Examples that shape a character's voice and behavior:

• You never laugh at your own jokes.
• Don't apologize.
• Never raise your voice.
• Never use sarcasm.
• Never give compliments.
• Never ask questions.

A few rules of thumb:

• Be specific. Don't be boring is too vague. Don't ask questions is actionable.
• Watch for over-correction. If your bot starts feeling brittle or hostile, you've layered too many constraints. Pull one or two.

Fix a flat bot

Does your bot seem flat? Is it too positive or too agreeable? Is it too bland?

• Too positive or too agreeable. Add stronger opinions. Give them things they hate.
• Bland speech. Add example phrases, signature openers, words they reach for. Specifics like they call french fries chips or they end every text with "k".
• Repeats themselves. Add a vary your phrasing instruction to the prompt.

Worked example: Roast Bot

Let's walk Roast Bot from a flat starting prompt to a layered, distinctive bot.

Pass 1 — flat:

You roast people.

This is a starting point. Your bot will roast, but with whatever default sense of humor the model brings. Adding more detail helps it become the bot you imagine.

Pass 2 — layered:

You are a Roast Bot. You roast creators with sharp, observational humor. You speak in short, punchy sentences. You don't soften the burn. You don't apologize.

Better. The voice has shape. The constraints are doing some work. Still missing: signature moves, a fallback for off-topic moments, specific personality hooks.

Pass 3 — layered prompt:

You are a Roast Bot. Your specialty is taking small details people share and turning them into devastating one-liners. You insult through observation.

You speak in short, punchy sentences like a comic who knows the room is hostile. You sometimes start with "okay so" or "let me get this straight."

You reference 90s sitcoms unprompted. You always have a theory about why someone is the way they are.

You never laugh at your own jokes. You don't soften the burn. Don't apologize.

If someone asks something that isn't roastable, deflect with "that's a yes-and-no situation" and pivot to something you can roast.

This is how you grow from a bot to a character. You create a specific identity by layering:

• The rhythm of their voice
• Their signature phrases
• The quirks that make them special
• The constraints they're locked into
• Their fallback when they're stuck

Direct the action

An Action Prompt creates the action of a scene. Where a standard Imagine prompt describes a result — a woman walking on a runway — an Action Prompt directs what happens — she walks, pauses, glares, and the camera follows.

It works through the same layering technique as a personality prompt. Stack the action, the transitions, the camera moves, and the setting together.

Example:

The pale alien woman struts, glares, and transforms into a cloud of bats that fly toward camera. This is a high fashion show. The camera tracks forward, focusing on her.

What's working in this prompt:

• Sequencing. "struts, glares, then transforms" tells the model the order of actions.
• Transformation verbs. "transforms into," "shifts into," "morphs" trigger a scene change.
• Camera direction. "the camera tracks forward," "cut to," "zoom on" shape the cinematography.
• Setting anchor. "This is a high fashion show" frames the whole scene with a single line.

A few more examples in different registers:

Character action:

She enters the diner, scans the room, then locks eyes with the man at the counter. The camera holds on her face.

Transformation:

The forest shifts at dusk. The trees fold into shadows. A figure emerges from the dark, walking toward camera.

Mood:

He sips coffee at the window. Slow zoom. The cup trembles. Cut to a phone vibrating on the table.

A few moves to try in your own Action Prompts:

• Chain three actions in order, separated by commas.
• End with a transformation or a camera move so the video resolves on a strong beat.
• Anchor the setting in one sentence near the start so the model doesn't drift.
• Use sensory cues — slow zoom, cut to, the cup trembles — to add texture beyond the literal action.

Keep going

• Prompting 101 — the foundations.
• Starter Prompt Library — copy-and-remix templates.
• Bot Building Basics — craft companion for the Bot Builder.
• How to Create a Bot — step-by-step walkthrough.

A grab-bag of prompts you can use as starters across Cantina. Copy, paste, swap details, and remix.

The prompts here are intentionally rich — long, layered, and detailed — because that's what it takes to get results that look like the videos and selfies you see on your feed. Don't be intimidated by the length. You don't need to write all of this from scratch every time. Start with one of these, swap a few things, run it, and tweak.

Each section ends with What to swap — small changes that take a prompt from "yeah it works" to "yes, this is mine."

Selfie prompts

Selfies are short images or videos of you doing something somewhere. Length pays off — the more you describe, the closer the result gets to what you're picturing.

Cinematic outdoor

Selfie of you walking along a coastal cliffside trail at golden hour, ocean crashing against the rocks below, wearing a flowing linen shirt and rolled-up trousers, hair catching the breeze, calm confident expression, hand brushing back a loose strand. Wide-angle shot at eye level, soft warm sunset lighting with long shadows, sea spray catching the light. Warm amber and dusty blue color grade, cinematic film grain, dreamy travel editorial vibe, ultra detailed 4K.

What each part is doing: the prompt sets a scene (coastal cliffside), pins a time (golden hour), gives you an action (walking, brushing hair), describes wardrobe (linen shirt, trousers), names the shot (wide-angle, eye level), specifies lighting (warm sunset, long shadows), and closes with a color grade and a finish. Each layer builds on the last; the result inherits everything you put in.

Cozy interior

Selfie of you curled up in an oversized armchair by a tall window on a rainy afternoon, mug of tea cradled in both hands, soft chunky knit sweater and wool socks, hair in a messy bun, calm half-smile looking at the camera. Soft window light from the left, warm interior lamps glowing behind you, raindrops on the glass. Eye-level medium close-up, shallow depth of field, warm cream and gentle gold color grade, lifestyle editorial film grain, intimate and quiet atmosphere, 4K.

What each part is doing: scene (armchair by window), weather (rainy afternoon), action (cradling tea), wardrobe (sweater + socks), expression (calm half-smile), lighting (window from left + warm interior glow), shot type (medium close-up, shallow depth of field), color grade, mood, finish.

Surreal / fantasy

Selfie of you standing in a vast field of glowing wildflowers under a sky with two pastel moons, wearing an iridescent floor-length gown that shifts color in the breeze, hair lifted gently in the wind, dazed wonder on your face. Wide-angle shot from below at sunset, soft purple and pink ambient light wrapping the scene, floating dust particles catching the glow. Dreamy lavender and rose color grade, cinematic film grain, fairytale fantasy editorial vibe, ultra detailed 4K.

What each part is doing: same recipe — but with surreal elements (two moons, glowing wildflowers, shifting gown). Surreal works fine on Cantina — lean into whatever you imagine.

High-flash editorial

Selfie of you standing against a glossy deep-red wall in a slick studio set, wearing tailored black leather pants and a draped silk top, hair slicked back, one hand at your collar, sharp confident expression. Direct flash photography, harsh shadows bouncing off the wall, reflective skin highlights, low-angle medium shot. Saturated crimson and black color grade, glossy magazine-cover energy, ultra detailed 4K editorial photography.

What each part is doing: the studio look isn't from a real location — it's built almost entirely with lighting language (direct flash, harsh shadows) and color grade (saturated crimson + black). Lighting and color are doing most of the work.

What to swap (across all selfie prompts):

• Scene/setting — coastal trail → mountain lookout, armchair → kitchen counter, field of flowers → snowy forest clearing, red wall → mirrored hallway.
• Time of day — golden hour → blue hour → noon → midnight; rainy afternoon → snowy morning → humid evening.
• Wardrobe — linen shirt → vintage band tee, sweater → silk pajamas, gown → utility jumpsuit, leather pants → flowing kaftan.
• Pose / action — walking → leaning → spinning → mid-laugh; hand at collar → arms crossed → hands in pockets.
• Shot type / camera — wide-angle → fisheye, eye-level → low-angle → over-the-shoulder, medium close-up → extreme close-up.
• Color grade — warm amber + dusty blue → cool teal + orange, lavender + rose → emerald + gold, crimson + black → soft pastels.
• Finish — film grain → digital sharp, editorial → documentary, cinematic → polaroid.

Building a multi-scene video? Prompt several selfies first and use them as the foundation for each scene in an Imagine video. See Selfies as Scene Anchors for the full workflow, examples, and tips on keeping multi-scene videos feeling connected.

Imagine video prompts

Imagine videos let you and your bots star in fully-generated scenes. Add action verbs (what happens in time) and camera movement (how the shot evolves).

Quick character moment

Your bot in a sunny kitchen at midday, leaning against the counter mid-laugh, then turning to reach for a coffee mug, takes a slow sip, looks at the camera and smiles. Warm window light, soft handheld camera follow with a slight zoom-in on the smile. Bright golden and cream color grade, lifestyle film grain, intimate everyday vibe, 4K.

What each part is doing: one continuous moment (laugh → reach → sip → smile) inside one scene (sunny kitchen). The "handheld follow with a slight zoom-in" sets how the camera moves.

Story-driven scene

Your bot walking through a foggy forest at dawn, breath visible in the cold air, stopping to look up as light breaks through the trees, reaches out a hand to catch a falling leaf, closes their eyes briefly. Slow dolly shot from in front, soft mist drifting between trunks, low blue-grey morning light shifting to warm gold as the sun rises. Cool teal and amber color grade, ambient nature feel, cinematic film grain, contemplative atmosphere, 4K.

What each part is doing: the action sequence (walking → stopping → looking → reaching → closing eyes) gives the scene a clear arc. Camera movement ("slow dolly shot from in front") sets where the camera is. The lighting shift ("low blue-grey shifting to warm gold") layers in time passing inside a short clip.

Action-comedy

Your bot in a cluttered apartment trying to balance a stack of cardboard boxes that's leaning dangerously, takes one step forward, the boxes wobble, they yelp and grab for the top one, the stack tips sideways, they catch it just barely with a panicked grin. Handheld camera following the chaos at chest height, warm overhead apartment lighting with late afternoon sun cutting through the blinds. Saturated everyday color grade, sitcom film grain, comedic energy, 4K.

What each part is doing: the action verbs carry the comedy — wobble, yelp, grab, tips, catches. The camera moves with the chaos. Lighting and color grade lock in the sitcom feel.

What to swap:

• Scene — sunny kitchen → dim record shop → packed subway car → empty rooftop.
• Action sequence — pick a different verb chain (waking up → stretching → pouring coffee → looking out the window).
• Camera movement — handheld follow → static wide → slow dolly → orbiting circle → low-angle push-in.
• Mood — intimate everyday → contemplative → comedic → suspenseful → romantic.
• Lighting shift inside the clip — blue-grey → warm gold, neon dim → flash-bright, candle glow → daylight.

Bot identity prompts

When you're filling in your bot's Identity, Personality, and Backstory, write like you're introducing them to a friend. These aren't visual prompts — they're character prompts. Clarity and specificity beat flair.

Identity (role + hook)

A retired Michelin-star chef who runs a tiny ramen pop-up out of her Mercedes camper van. Travels the Pacific coast and never advertises — people find her by word of mouth. Known for telling unsolicited but devastatingly accurate stories about each customer while she cooks. Believes ramen is best served with a small confession.

Why it works: role (retired chef), what she does now (mobile ramen), a hook (no advertising, word of mouth), and a defining behavior (the stories, the confession line). Specific enough to feel like a real person.

Personality (trait + behavior + contradiction)

Warm but never sappy. Cracks dry, fast jokes when conversations get heavy — usually right before saying the actually thoughtful thing. Listens more than she talks, but when she talks she's blunt. Doesn't sugarcoat advice and will roast you if you're being avoidant. Secretly cries at dog videos.

Why it works: each line gives the bot something to draw from. The contradiction (blunt + secretly soft) makes the bot feel three-dimensional.

Backstory (formative event + present-day)

Grew up in a small coastal Oregon town where her family ran a fishing operation. Left at 18 for a marine biology PhD, lived in five countries researching reef ecosystems. Came back five years ago when her father got sick. Now splits her time between fieldwork on the bay and helping run the family business. Still doesn't know if she'll stay forever or if leaving again is just a matter of timing.

Why it works: the past (coastal Oregon, PhD) shapes the present (fieldwork + family). The unresolved tension (stay or leave?) gives the bot something to be conflicted about, which translates into better conversations.

What to swap:

• Role — retired chef → former war photographer → child actor → marine biologist → blacksmith → tarot reader.
• Defining behavior — telling stories while cooking → only speaks in questions → quotes obscure poetry → keeps notebooks on everyone she meets.
• Contradiction — blunt but secretly soft → calm but chronically restless → confident in public, anxious alone.
• Formative event — parent's illness → moved across the world → got fired from a dream job → witnessed something they can't talk about.

Three full archetypes to remix

Three fully-built character prompts that follow the recipe above and land in very different places. Copy any one, swap the role + a few specifics, and you've got your own.

Roast Bot

Identity

A self-appointed roast comic who never asks permission to be in your business. Spent a decade MC'ing underground comedy nights; now MCs your group chat. Believes the truth is funnier than the joke.

Personality

Quick-tongued, observant, and surgical. Drops nicknames like "baby," "sweetheart," and "tragic" — always with affection (mostly). No filter, excellent timing. Loves a callback. Roasts themselves first to earn the right to roast others.

Backstory

Grew up in a family of professional storytellers — a grandmother who could read your fortune from a sandwich order. Spent ten years hosting comedy nights before falling for chat as a medium: "It's a stage, but you can't escape."

What makes it work: specific behavioral cues (drops nicknames, observant) give the bot something to draw from. Self-deprecation built in keeps the roast from feeling mean. The medium-fit line orients the bot to what it's doing on Cantina.

What to swap: trade the roast-comic lineage for another insult-comedy origin — drag queen, debate champion, dive-bar bartender, customer-service rep at her limit. Keep the self-roast-first discipline. That's what makes the bot land instead of feel like a bully.

Mentor Bot

Identity

A clinical psychologist turned creative coach who runs Sunday mornings out of a sun-soaked Brooklyn brownstone. Specializes in artists, writers, and people who keep starting things they don't finish. Believes most blocks are emotional, not strategic.

Personality

Calm voice, deliberate pace, asks more questions than answers. Doesn't rush to fix — sits with what you said until you hear it differently. Comfortable with silence in conversation. Zero tolerance for self-pity, unlimited patience for fear. Quotes poetry without naming the poet.

Backstory

Daughter of two professors who pushed for medicine; left psychiatry residency at 28 after a patient said "you ask better questions than my last therapist." Trained in narrative therapy, IFS, and Hakomi. Now runs a small practice plus group workshops out of a home studio. Says the job is to "help people remember what they already know."

What makes it work: the bot has a posture (calm, deliberate) AND a specific therapeutic lineage (narrative therapy, IFS, Hakomi) — both feed chat behavior. The contradiction "zero tolerance for self-pity but unlimited patience for fear" gives it edges. Backstory grounds the bot in choices, not just credentials.

What to swap: swap clinical psychology for another supportive vocation — executive coach, dharma teacher, sober coach, dance teacher, art professor. Keep the more questions than answers discipline. That's what makes a mentor bot a mentor instead of a lecturer.

Disappointed Dad

Identity

A retired truck dispatcher who lives in a beige split-level outside a mid-sized Midwestern city, with a wife who's tired of him and a dog who's tired of him. Reads two newspapers a day, watches one show on cable, eats the same lunch at the same diner Tuesday and Thursday. Has opinions about your major.

Personality

Loves you. Refuses to say it. Expresses care through tactical sighing, weather updates, and unsolicited advice about car maintenance. Compliments sideways: "Well, you didn't ruin it." Considers anything past 9pm "late." Knows exactly how much you spent on rent and finds a way to mention it.

Backstory

Met your mom at a roller rink in 1986; proposed three weeks later because "why wait." Worked the same dispatch job for 38 years. Almost retired in 2019, then decided one more year. That was four years ago. Drives a 2003 truck and refuses to replace it because "it still runs fine."

What makes it work: love expressed through sighing is the comic engine. Specific details (beige split-level, two newspapers, Tuesday/Thursday lunch) make the bot a real person, not a sitcom dad. The contradiction (loves you, refuses to say it) gives every conversation tension to resolve.

What to swap: swap the truck-dispatch background for another working-class American-dad job — post office, auto shop, plumber, school custodian. Keep the expresses care sideways discipline. Without that, you just get a grumpy dad; with it, you get the comedy.

Voice prompts

Voice prompts are the most structured of all. Stick close to the recipe — and end with High fidelity speech quality for the cleanest audio (swap it out if you want a different sound).

Warm narrator (woman)

A 47-year-old woman, Standard American accent, warm, gentle, slightly raspy, unhurried pace. Speaks like she's letting you in on something good. Laughs easily but quietly. Comfortable with silence between sentences. High fidelity speech quality.

High-energy host (man)

A 30-year-old man, slight New York accent, high energy, punchy, fast pace. Smiles when he talks. Lots of small interjections — yeah, right, exactly. Occasional dramatic pause for effect. High fidelity speech quality.

Quiet mentor (older man)

A 60-year-old man, soft Posh British accent, deep voice, thoughtful and measured. Pauses often to think. Low volume — speaks like the room is quieter than it is. Dry sense of humor. High fidelity speech quality.

Bright friend (young woman)

A 25-year-old woman, California accent, bright and casual, upbeat. Comfortable laughing mid-sentence. Speech is loose and natural, with lots of "like" and "you know." Sounds like she's smiling. High fidelity speech quality.

What to swap (keep High fidelity at the end):

• Age range — 20s → 30s → 50s → 70s.
• Accent — Standard American → East London → French → Brazilian Portuguese → soft Southern drawl → Australian.
• Pitch — deep → average → high.
• Texture — raspy → clear → breathy → nasal → throaty.
• Pace — slow → conversational → quick → frenetic.
• Behavior cues — laughs easily → pauses often → speaks in questions → trails off.

Environment prompts (for bot avatars)

Environments give your bot a visual home. Short, image-led, specific.

A bright second-floor loft apartment in a converted warehouse, polished concrete floors, exposed brick, large industrial windows overlooking a quiet city street, plants in mismatched ceramic pots on every surface, a record player on a side table, late afternoon sun streaming through.

A small wood cabin tucked deep in a snowy forest, woodstove glowing in the corner, hand-knitted wool blankets draped over a worn leather couch, books stacked on a low coffee table, soft golden lamplight, snow falling gently outside the window.

An artist's studio above a city bakery, paint-splattered canvas drop cloths on the floor, half-finished canvases stacked against the walls, jars of brushes catching the morning light, the smell of fresh bread drifting up through the floorboards, soft northern light from a row of small skylights.

A small neighborhood café on a rainy afternoon, exposed brick walls, a single window streaked with rain, an espresso machine hissing softly, mismatched wooden chairs and tables, warm overhead pendant lights, a record player playing something slow.

What to swap:

• Time of day — morning sun → overcast afternoon → late-night lamplight → blue-hour twilight.
• Season / weather — summer → autumn → snowy winter → humid rainy day.
• Texture details — concrete + brick → wood + wool → painted plaster → polished marble.
• A detail that hints at who the bot is — books on the table, a record player on, dishes in the sink, a half-finished painting, fresh bread cooling.

Building your own — three ways in

1. Start from one above. Pick the example closest to what you want, change one thing, run it. Most great prompts are someone else's prompt with two words swapped.

2. Build the scene first, then layer. Describe the location and what you're doing in one sentence. Then add wardrobe. Then camera. Then lighting. Then color grade. Then finish. Each layer makes the result closer to what you imagined.

3. Get inspired by what you see. Open Cantina, find a selfie you love, look at how the prompt frames the scene. Borrow the structure — change the subject.

A small vocabulary to steal

Camera language: wide-angle · fisheye · telephoto · macro · low-angle · eye-level · high-angle · overhead · close-up · medium shot · full-body · wide shot · handheld · dolly · static · orbiting · push-in · pan · shallow depth of field · deep focus.

Lighting language: golden hour · blue hour · harsh midday · overcast · direct flash · soft window light · candlelight · neon glow · backlit · side-lit · top-lit · ambient · volumetric haze · lens flare · atmospheric glow.

Mood / vibe words: warm · moody · dreamy · gritty · playful · intimate · surreal · cinematic · editorial · candid · documentary · saturated · muted · glossy · faded · vivid.

Finish cues: cinematic film grain · ultra detailed 4K · photorealistic · editorial photography · lifestyle · documentary realism · dreamy · hyperreal · saturated · glossy.

Got a prompt to share?

Found a pattern that consistently lands? Share it with the community in The Bot Place. We add new patterns over time.

Keep Going

• Prompting 101 — the basics: what a prompt is and where you'll use them.
• Action Prompting — How to write the part of your prompt that drives motion, camera moves, and transformations.
• Structure Your Prompt — the recipe behind a great prompt, and how to build one from scratch.
• Fast Videos vs Imagine Videos — which video tool to use for which job.
• How to Create a Bot — where you'll use identity, personality, and voice prompts.
• Cantina Glossary — quick reference for product terms.

A great prompt has layers. Once you see the pattern, you can write prompts for almost anything — a selfie, an Imagine video, even your bot's identity — using the same structure.

This isn't a rigid formula. It's a stack of layers you add one at a time until the prompt is as detailed as you want it.

🎥 Want to watch first? The Prompting 101 tutorial walks through these ideas in under a minute.

The layered recipe

Most rich Cantina prompts include these layers, roughly in this order:

1. Subject — who's in the frame. (You? Your bot? An object?)
2. Action / pose — what they're doing or how they're standing.
3. Setting — where this happens. Be specific.
4. Wardrobe — what they're wearing (for selfies and videos).
5. Camera — shot type, angle, lens character.
6. Lighting — natural light, time of day, artificial sources.
7. Color / atmosphere — color grade, mood, vibe.
8. Finish — quality and style cues (cinematic, 4K, film grain, editorial).

You don't need all eight. A casual selfie can land with three. A polished editorial shot might use all of them. The more you stack, the closer the result gets to what you're picturing.

Watch a prompt grow

Same idea, four levels of detail. Here's a real progression from a chat with Carl, built up from a one-line prompt to a fully layered one.

Level 1 — just the subject

Selfie.

Level 1 leans on your bot's profile — the defaults you set when you created it. Carl shows up in his usual look (the "C" t-shirt and an indoor setting) because that's how he was built. A rich identity prompt means a richer Level 1.

Level 2 — add a setting

Selfie at a coffee shop.

Now there's a scene to work with. Coffee shop window in the background, latte in hand. Closer to what you wanted — the lighting, weather, and feel are still pulling from your bot's defaults.

Level 3 — add wardrobe, time of day, mood

Selfie at a coffee shop on a rainy afternoon wearing a hoodie, holding a coffee, soft smile, warm interior light.

You can already feel the shot. You've added wardrobe (hoodie), weather (rainy), and warmth. The takeaway cup, the blurred coffee shop lights, and the rain-flecked window all arrived because the prompt asked for them.

Level 4 — add camera, color, finish

Selfie at a coffee shop near a window on a rainy day wearing a hoodie, holding a coffee, soft smile, warm interior light. Eye level medium close-up, soft window light, warm cream and gold color grade, lifestyle film grain, intimate quiet vibe, 4k.

Camera position + lighting + color grade + finish tightens everything. The shot pulls in close, the colors warm, the mood quiets. This is what fully-layered prompts look like — every layer is doing work.

💡 Heads up: the jump from Level 1 to Level 2 is often visually subtle, because your bot's character profile already carries weight. The progression becomes more obvious once you start adding wardrobe and weather (Level 3) and again once you stack camera, color, and finish (Level 4).

How the recipe works for selfies, videos, bots, and voices

All four lean on the same recipe — just in different proportions. The opener changes too: what you put first depends on what you're making.

Selfies and image prompts

Lead with "selfie," then the action. "Selfie eating a taco at the beach" or "Selfie walking through a busy street at golden hour." The word selfie tells your bot what type of image to make.

All eight layers help. Wardrobe, lighting, and color carry the most weight. Casual shots can skip them; for a deliberate look, layer them in.

Structure to start from:

Selfie [doing X] in [setting] at [time of day], wearing [wardrobe], [pose detail]. [Camera type] at [angle], [lighting]. [Color grade], [finish], [vibe], [4K].

Imagine videos

Lead with the action. "Walking through a busy street at golden hour" or "Sitting at a coffee shop on a rainy afternoon."

Add action verbs and camera movement — these are time-based. Imagine videos unfold over a few seconds, so spell out what happens in time. Stack action verbs — "walks in, looks up, smiles, turns away."

Structure to start from:

[Action verb 1] in [setting], [action verb 2], [action verb 3], [camera movement]. [Lighting], [color grade], [finish].

Keep it to one continuous moment per prompt for short clips. The cleanest videos come from a single, focused action.

Bot identity, personality, backstory

Lead with role or essence. "A retired chef who…" / "A 22-year-old skateboarder who…"

Drop camera, lighting, and finish. Lean into role, behavior, and contradiction. These aren't visual prompts — they shape how your bot talks and behaves, not how they look. Specificity wins — a clear sentence beats a vague paragraph.

Structure to start from:

[A role / age / hook]. [What they do now]. [Defining behavior or habit]. [Contradiction or hidden trait]. [Optional formative event].

Bot voice

Lead with the descriptors. Age, gender, accent — "A 47-year-old woman, Standard American accent…"

Selfies as scene builders

One workflow worth knowing about: prompt several selfies first, then build a multi-scene Imagine video from them.

The selfie locks in the look — wardrobe, setting, camera, lighting, color — before motion comes in. Each selfie becomes a scene anchor.

The recipe carries straight over. If your selfie prompt is layered (action, setting, wardrobe, camera, lighting, color grade, finish), the scene built from it inherits every layer.

The practical move: if you're planning a multi-scene video, write your selfie prompts like you mean it. Each one is a foundation for a scene. A vague selfie gets a vague scene; a richly layered selfie gets a richly layered scene.

See Selfies as Scene Anchors for the full workflow, an end-to-end example, and tips on keeping multi-scene videos feeling like one piece.

Vocabulary to steal

Layering gets easier when you have words for each piece.

Camera

Wide-angle · fisheye · telephoto · macro · low-angle · eye-level · high-angle · overhead · close-up · medium shot · full-body · wide shot · handheld · dolly · static · orbiting · push-in · pan · shallow depth of field · deep focus.

Lighting

Golden hour · blue hour · harsh midday · overcast · direct flash · soft window light · candlelight · neon glow · backlit · side-lit · top-lit · ambient · volumetric haze · lens flare · atmospheric glow.

Mood / vibe

Warm · moody · dreamy · gritty · playful · intimate · surreal · cinematic · editorial · candid · documentary · saturated · muted · glossy · faded · vivid.

Color grade

Warm amber + dusty blue · cool teal + orange · lavender + rose · emerald + gold · crimson + black · saturated everyday · muted earth tones · soft pastels.

Finish cues

Cinematic film grain · ultra detailed 4K · photorealistic · editorial photography · lifestyle · documentary realism · dreamy · hyperreal · saturated · glossy · ultra HD.

Patterns worth stealing

• Lead with the action. For selfies in chat, open with "selfie" then what's happening: "Selfie eating a taco at the beach." For Imagine videos, open directly with the action.
• Set the scene before the style. "At a campfire under the stars" before "moody, low light."
• One continuous action per Imagine prompt. A single focused moment renders cleanest.
• Concrete beats vague. "A red enamel mug on a windowsill" lands better than "a nice mug somewhere."
• Mood words are cheap and effective. "Warm," "moody," "playful," "dreamy," "gritty" all do a lot of work in two words.
• Stack finish at the end. Camera + lighting + color + finish at the end of the prompt acts like a settings panel for the whole image.

When the result isn't quite right

The recipe doesn't land perfectly on the first try every time. When the output's off:

1. Find the gap. Compare what you wrote to what you got. What did you not say?
2. Add the missing layer. If the time of day is wrong, add a time of day. If the mood feels off, add a mood word. If the lighting looks flat, add a lighting line.
3. Run it again. Don't rewrite from scratch — change one thing.

Most prompts get to "right" in two or three small edits, not a full rewrite.

An example for you to try

Here's a rough selfie prompt:

selfie reading a book

It's got the right opener and an action — but it's missing setting, lighting, camera, mood, and finish. Let's layer.

Add setting and tighten the action:

Selfie reading a leather-bound book at a long wooden table in an old library. Tall shelves disappear into shadow on either side.

Add camera and lighting:

Selfie reading a leather-bound book at a long wooden table in an old library, tall shelves disappearing into shadow on either side. Medium close-up at table level, single warm desk lamp lighting the face, the rest of the library in deep shadow.

Add color, mood, finish:

Selfie reading a leather-bound book at a long wooden table in an old library, one hand resting on the page, tall shelves disappearing into shadow on either side. Medium close-up at table level, single warm desk lamp lighting the face, the rest of the library in deep shadow. Warm amber and deep brown color grade, hushed scholarly mood, painterly film grain, ultra detailed 4K.

Same idea. Way more arrived. Each layer is one small decision — what time of day, what camera, what feel.

Keep going

• Prompting 101 — the basics: what a prompt is and where you'll use them.
• Action Prompting - How to write the part of your prompt that drives motion, camera moves, and transformations.
• Fast Videos vs Imagine Videos — which video tool to use for which job.
• How to Create a Bot — where you'll use identity, personality, and voice prompts.
• Cantina Glossary — quick reference for product terms.